Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models

Table 6 The performance comparison (average RMSE) of the 50 times independent runs on the three regression datasets for the eight models. (the top three model were italic for each dataset)

Dataset	No.	Tasks	Metric	Model	Training	Validation	Test
ESOL	1127	1	RMSE	SVM	0.149 ± 0.005	0.565 ± 0.038	0.569 ± 0.052
				XGBoost	0.224 ± 0.057	0.573 ± 0.048	0.582 ± 0.056
				RF	0.391 ± 0.008	0.664 ± 0.053	0.663 ± 0.074
				DNN	0.492 ± 0.061	0.617 ± 0.060	0.670 ± 0.092
				GCN	0.272 ± 0.049	0.650 ± 0.064	0.708 ± 0.068
				GAT	0.300 ± 0.093	0.608 ± 0.083	0.658 ± 0.109
				MPNN	0.463 ± 0.074	0.652 ± 0.051	0.700 ± 0.073
				Attentive FP	0.390 ± 0.076	0.535 ± 0.045	0.587 ± 0.065
FreeSolv	639	1	RMSE	SVM	0.307 ± 0.023	0.804 ± 0.192	0.852 ± 0.171
				XGBoost	0.228 ± 0.168	0.988 ± 0.197	1.025 ± 0.185
				RF	0.518 ± 0.011	1.129 ± 0.248	1.143 ± 0.230
				DNN	0.574 ± 0.115	0.840 ± 0.158	1.013 ± 0.197
				GCN	0.703 ± 0.127	0.872 ± 0.191	1.149 ± 0.262
				GAT	0.937 ± 0.375	1.079 ± 0.204	1.304 ± 0.272
				MPNN	0.824 ± 0.220	1.130 ± 0.245	1.327 ± 0.279
				Attentive FP	0.720 ± 0.131	0.881 ± 0.207	1.091 ± 0.191
Lipop	4200	1	RMSE	SVM	0.191 ± 0.005	0.566 ± 0.037	0.577 ± 0.039
				XGBoost	0.191 ± 0.040	0.569 ± 0.033	0.574 ± 0.034
				RF	0.478 ± 0.003	0.660 ± 0.031	0.659 ± 0.031
				DNN	0.271 ± 0.068	0.583 ± 0.031	0.608 ± 0.034
				GCN	0.360 ± 0.081	0.616 ± 0.038	0.664 ± 0.086
				GAT	0.372 ± 0.084	0.658 ± 0.037	0.683 ± 0.060
				MPNN	0.476 ± 0.065	0.640 ± 0.037	0.673 ± 0.038
				Attentive FP	0.309 ± 0.045	0.533 ± 0.033	0.553 ± 0.035

ISSN: 1758-2946